Hyperspectral sensing for turbid water quality monitoring in freshwater rivers: Empirical relationship between reflectance and turbidity and total solids.

Total suspended solid (TSS) is an important water quality parameter. This study was conducted to test the feasibility of the band combination of hyperspectral sensing for inland turbid water monitoring in Taiwan. The field spectral reflectance in the Wu river basin of Taiwan was measured with a spectroradiometer; the water samples were collected from the different sites of the Wu river basin and some water quality parameters were analyzed on the sites (in situ) as well as brought to the laboratory for further analysis. To obtain the data set for this study, 160 in situ sample observations were carried out during campaigns from August to December, 2005. The water quality results were correlated with the reflectivity to determine the spectral characteristics and their relationship with turbidity and TSS. Furthermore, multiple-regression (MR) and artificial neural network (ANN) were used to model the transformation function between TSS concentration and turbidity levels of stream water, and the radiance measured by the spectroradiometer. The value of the turbidity and TSS correlation coefficient was 0.766, which implies that turbidity is significantly related to TSS in the Wu river basin. The results indicated that TSS and turbidity are positively correlated in a significant way across the entire spectrum, when TSS concentration and turbidity levels were under 800 mg·L(-1) and 600 NTU, respectively. Optimal wavelengths for the measurements of TSS and turbidity are found in the 700 and 900 nm range, respectively. Based on the results, better accuracy was obtained only when the ranges of turbidity and TSS concentration were less than 800 mg·L(-1) and less than 600 NTU, respectively and used rather than using whole dataset (R(2) = 0.93 versus 0.88 for turbidity and R(2) = 0.83 versus 0.58 for TSS). On the other hand, the ANN approach can improve the TSS retrieval using MR. The accuracy of TSS estimation applying ANN (R(2) = 0.66) was better than with the MR approach (R(2) = 0.58), as expected due to the nonlinear nature of the transformation model.